ArcFace / FaceNet
Definition
Two influential deep CNN facial recognition architectures. FaceNet (Google, 2015) used a triplet-loss training approach to learn facial embeddings. ArcFace (Deng et al., 2019) introduced an additive angular margin penalty that improved embedding discrimination and has become a widely adopted reference architecture.
- FaceNet developer
- Google, 2015
- FaceNet method
- Triplet-loss training for facial embeddings
- ArcFace developer
- Deng et al., 2019
- ArcFace method
- Additive angular margin penalty
Common questions
What problem does ArcFace's angular margin penalty solve that FaceNet's triplet loss did not fully address?+
The angular margin pushes embeddings of the same identity closer together and different identities further apart in angular space more directly than triplet loss, which depends heavily on how training triplets are selected, giving ArcFace more consistent discrimination across large datasets.
Why do these architectures matter to forensic facial recognition rather than just consumer applications?+
Both produce the embedding vectors that most modern one-to-many facial recognition search systems used in investigations rely on, so understanding their training approach and known failure modes, such as sensitivity to pose, lighting, and demographic bias, is relevant when an examiner assesses a system's output.
Related terms
- Embedding
- A compact numerical vector representation of a face produced by a neural network. The distance between two embeddings in the vector space...
- False Match Rate (FMR)
- The proportion of non-mate comparisons (different individuals) that the system incorrectly scores as a match. Also called the false positive rate or...
- False Non-Match Rate (FNMR)
- The proportion of mate comparisons (same individual) that the system incorrectly scores as non-matching. Also called the false negative rate. A lower...
- Gallery vs Probe
- In facial recognition, the gallery is the reference database of enrolled face images; the probe is the query image being compared against...
- NIST FRVT
- The National Institute of Standards and Technology Face Recognition Vendor Testing programme, the primary independent benchmark for facial recognition algorithms, publishing ongoing...